Anthropic Exposes “Industrial-Scale” Distillation Attacks on Claude From China
Anthropic said the attackers created roughly 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude
Anthropic, the U.S. AI developer behind the Claude large language model, has disclosed a series of large-scale “distillation attacks” on its systems, highlighting risks to model security and broader AI competitiveness.
In a blog post published on February 23, the company identified coordinated campaigns by three AI laboratories — DeepSeek, Moonshot AI, and MiniMax — that illicitly extracted Claude’s capabilities to train or improve their own models.
Last year, OpenAI significantly tightened its security protocols, particularly after concerns that Chinese AI firm DeepSeek may have copied its models using distillation techniques.
Anthropic said the attackers created roughly 24,000 fraudulent accounts and generated more than 16 million exchanges with Claude in violation of its terms of service and regional access limits.
“We have identified industrial-scale campaigns by three AI laboratories… to illicitly extract Claude’s capabilities to improve their own models,” the company wrote, describing how distillation — normally a legitimate AI training method — was abused at scale.
The blog warns that models created through such unauthorized distillation lack essential safety protections, posing “significant national security risks” if deployed without safeguards. Anthropic argues that these campaigns undermine efforts like U.S. export controls on advanced AI chips by enabling labs with restricted access to extract frontier capabilities indirectly.
To counter the threat, Anthropic said it is strengthening detection tools, behavioural fingerprinting systems, stricter account verification, and sharing intelligence with cloud providers and other AI developers.
It also urged “rapid, coordinated action among industry players, policymakers, and the global AI community” to address the growing sophistication of such attacks.
Last week, the startup introduced Claude Sonnet 4.6, the latest version of its mid-tier AI model, touting major gains in coding, long-context reasoning, computer use and agentic planning as competition among frontier AI systems intensifies.
